Moral Concern Is a Response Regime: A Preregistered Cross-Model Test of Persona Framing in Large Language Models

<p>Four models from four providers, Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and DeepSeek-V4-Flash, judged the same 20 ambiguous scenarios under a neutral frame and six loaded personas across 11,200 main trials. The study tests whether persona framing moves large language model outputs in ways predicted by Motivated Violation Construction, a theory of human moral cognition in which actor identity can shift ambiguous behaviour into the violation category. Across the retained three-model confirmatory panel, no MVC prediction was supported. Two models classified almost every ambiguous case as concerning under the neutral frame, leaving little room for the predicted increase, and one of those failed the registered test-retest stability gate. Persona effects were heterogeneous, often reversed direction, and were sensitive to model-specific baseline, headroom, prompt position, and response stability. A registered narrative-coding arm failed its human-reliability gate and was excluded in full. On semantically inert stimuli, one model expressed concern on roughly a third of trials while the others almost never did so. The results do not support transferring MVC’s human behavioural signature to the models tested. More generally, they show that the same forced-choice prompt did not function as the same measurement instrument across systems. We describe this configuration as a model’s response regime: the observable pattern of baseline tendency, available headroom, retest behaviour, and sensitivity to framing within which a model’s forced verdicts are produced.</p>